A citation cannot do a quotation mark’s job
Anthropic’s proposed fix is to give Claude a complete example of a correctly written response, including the request and an explanation of why the answer is correct. Its example demonstrates paraphrasing and a clearly marked short quotation.
A source link tells readers where information came from. Quotation marks identify words taken directly from that source. A response can therefore point to the right document while still leaving readers unclear about whose wording they are reading.
Anthropic offers a separate citations feature for developers that connects responses to passages in supplied documents. It can identify page locations in PDFs and character positions in plain text. The company says this produces valid references to the supplied material and selects relevant supporting passages more reliably than asking for citations through a prompt alone.
That is useful infrastructure for checking an answer. It does not remove the need to compare the draft’s wording with its sources. The Fable 5.1 guidance provides no frequency estimate for the quotation issue, so it cannot establish how often editors will encounter it.
Lower effort can mean fewer searches
The Fable 5.1 guide also says the model searches less often than Fable 5 at low effort and is more likely to answer from memory. Anthropic suggests raising effort for affected turns or explicitly prompting verification of unfamiliar or rapidly changing names.
Effort is a setting developers can use to influence how much work Claude does on a request. Anthropic’s API documentation explains that it affects reasoning, responses and tool calls. Lower settings can improve speed and reduce token use, with a possible reduction in capability.
For an article comparing software, the practical distinction is straightforward: recognizing a product’s name does not establish its current features, pricing or availability. Those details need current evidence.
As SEW’s query fan-out explainer describes, AI research can involve several related searches. Having that capability available does not establish that it was used for a particular answer.
A cheaper draft can still be expensive to publish
Token savings measure the cost of running the model. A publishing team also pays for the work needed to make its output usable. If an editor has to reopen sources, identify copied sentences and repeat research the model skipped, that time belongs in the calculation. A lower API bill deserves limited applause if the missing work simply lands on someone else’s desk.
Anthropic’s suggested examples and citation tools can help. Its separate hallucination guidance also recommends grounding claims in supporting passages and allowing the model to admit uncertainty, while acknowledging that errors remain possible. These measures can make an answer easier to inspect. They do not establish that editorial review can be removed.
For publishers, the meaningful comparison is cost per publishable draft. Run the same assignments with each prompt and effort setting, then record factual corrections, unmarked copied passages and editing time alongside model costs. That applies Anthropic’s own evaluation guidance, which calls for task-specific measures covering quality, speed and price. The savings become credible when the complete job gets cheaper.
Start the conversation by posting the first comment